Heterogeneous data-driven marketing channel clustering modeling and optimizing method

Abstract

The invention discloses a marketing channel clustering modeling and optimizing method driven by heterogeneous data. The method comprises the steps of obtaining a multi-dimensional data set of a marketing channel, and performing preprocessing; the data set comprises structured data and unstructured data; and extracting a text word segmentation result and an emotion feature from the unstructured data, and extracting an image feature to generate a structured feature set, an unstructured feature set and the like. According to the method, structured and unstructured data can be integrated, and potential information of marketing channels can be comprehensively mined. Through a deep learning model (an auto-encoder, a variational auto-encoder and a generative adversarial network), features are extracted and optimized layer by layer, and the quality of the features and the performance of the model are improved. And in combination with a clustering algorithm and an evaluation index, the accuracy and reliability of a clustering result are ensured, and powerful support is provided for optimization of a marketing channel.

Description

Technical Field

[0001] The present invention relates to the technical field of marketing channel analysis, and in particular to a marketing channel clustering modeling and optimization method driven by heterogeneous data. Background Art

[0002] With the rapid development of e-commerce, social media, and mobile advertising, marketing channels are becoming increasingly diversified. Enterprises engage with users through data-driven channels such as online advertising, social platforms, search engines, and mobile apps. These channels exhibit significant differences in user behavior and conversion paths (average click-through rates fluctuate by 42% and conversion rates vary by over 300%). Multi-dimensional data analysis reveals that traditional single-channel optimization methods are no longer able to meet the needs of precision marketing, with return on investment (ROI) varying by as much as 5.8 times across different platforms. Through heterogeneous data modeling, enterprises can establish a channel effectiveness evaluation system, enabling data-driven, intelligent dynamic resource allocation and significantly improving the efficiency of marketing budget allocation. Existing marketing channel optimization research has largely focused on a single channel, lacking comprehensive analysis of cross-channel data. Traditional methods primarily rely on structured data and fail to fully exploit the potential for valuable marketing information in unstructured data (such as text, images, and videos). Furthermore, classic clustering algorithms (such as K-Means) are primarily based on shallow learning models and struggle to effectively handle the complex nonlinear relationships in high-dimensional, heterogeneous data, resulting in limited clustering effectiveness. As the scale and dimensions of marketing data continue to grow, how to use deep learning technology to improve clustering quality has become an important research topic in the current field of precision marketing. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to propose a heterogeneous data-driven marketing channel clustering modeling and optimization method, which can solve at least one technical problem mentioned in the background technology.

[0004] According to one aspect of the present invention, a heterogeneous data-driven marketing channel clustering modeling and optimization method is provided, the method comprising:

[0005] Acquire a multi-dimensional data set of marketing channels and perform preprocessing; the data set includes structured data and unstructured data;

[0006] Extracting text segmentation results and sentiment features from the unstructured data, and extracting image features to generate a structured feature set and an unstructured feature set;

[0007] Extracting efficiency features based on the structured feature set, combining them with sentiment features in the unstructured feature set, analyzing overlapping patterns of cross-channel user behaviors to generate a channel collaborative correlation feature set;

[0008] Construct a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network. The channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set.

[0009] Based on a deep comprehensive feature set, a clustering algorithm is used to divide marketing channels into different categories, generate preliminary classification results of marketing channels, evaluate the clustering quality through an evaluation index, and determine the optimal number of clusters to output the channel category attribution results.

[0010] In the aforementioned technical solution, the method aims to achieve accurate classification and optimization of marketing channels by integrating multi-dimensional data (including structured and unstructured data) from marketing channels, extracting key features, and building a deep clustering model. This process fully utilizes advanced data processing and machine learning technologies, providing data support for marketing strategy formulation and channel optimization.

[0011] Acquiring multi-dimensional data sets: Collecting various data related to marketing channels, including structured data (such as sales data and user demographics) and unstructured data (such as text reviews and images). The collected data is cleaned, formatted, and preliminarily processed to prepare for subsequent feature extraction and modeling.

[0012] Feature extraction: Perform word segmentation on text data to extract keywords and sentiment features from the text; perform feature extraction on image data to convert it into a structured feature set. Integrate text segmentation results, sentiment features, and image features to form a structured feature set and an unstructured feature set.

[0013] Cross-channel user behavior analysis: By analyzing the overlap of user behaviors across different marketing channels, we can identify user interaction patterns across multiple channels. Based on this user behavior analysis, we generate a feature set that reflects the synergy between channels, providing higher-level features for subsequent modeling.

[0014] Deep Clustering Model Construction: An autoencoder is used to extract core features from the channel-coordinated feature set, reducing data dimensionality while retaining key information. This core feature is then fed into a variational autoencoder to extract the data's distributional characteristics, providing a foundation for a generative adversarial network (GAN). This GAN further optimizes the feature extraction results, generating a deep, comprehensive feature set that enhances both expressiveness and discriminability.

[0015] Clustering and Evaluation: Based on a deep, comprehensive feature set, we use clustering algorithms (such as K-means and DBSCAN) to classify marketing channels into different categories and obtain preliminary classification results. We use evaluation indices (such as the silhouette coefficient and Calinski-Harabasz index) to assess clustering quality, determine the optimal number of clusters, and output the final channel classification results.

[0016] This method integrates structured and unstructured data to comprehensively mine potential insights from marketing channels. Using deep learning models (autoencoders, variational autoencoders, and generative adversarial networks), it extracts and optimizes features layer by layer, improving both feature quality and model performance. Combining clustering algorithms with evaluation indices ensures the accuracy and reliability of clustering results, providing strong support for optimizing marketing channels.

[0017] In some embodiments, extracting efficiency features based on the structured feature set, combining them with sentiment features in the unstructured feature set, and analyzing overlapping patterns of cross-channel user behaviors to generate a channel collaborative correlation feature set includes:

[0018] The channel collaborative correlation feature set is spliced ​​to form a multimodal feature matrix, and a variational autoencoder is used to reduce the dimensionality of the multimodal feature matrix to generate a 128-dimensional latent space representation;

[0019] Build cross-channel access sequences based on user behavior timestamps, use LSTM networks to model user path patterns, calculate conversion funnel indicators for each channel, and generate structured efficiency features;

[0020] If the sentiment feature shows that negative reviews exceed the threshold, a channel quality warning mark will be triggered;

[0021] Analyze user overlapping access patterns through graph convolutional networks and construct a channel correlation matrix;

[0022] Combining the latent space representation with the channel association matrix, a spectral clustering algorithm is used to generate channel co-groupings;

[0023] Based on the grouping results, the conversion path complementarity coefficients between channels in the group are calculated, and a three-dimensional feature tensor containing efficiency features, sentiment features and conversion path complementarity coefficients is output as the final channel collaborative correlation feature set.

[0024] In the aforementioned technical solution, the method aims to integrate structured and unstructured features. Through steps such as constructing and reducing a multimodal feature matrix, modeling user behavior sequences, performing sentiment analysis and early warning, and analyzing channel associations, a set of channel synergy-related features is generated. This provides a quantitative basis for a deeper understanding of inter-channel relationships and synergies, helping to optimize marketing strategies and resource allocation.

[0025] Multimodal feature matrix construction and dimensionality reduction: This involves concatenating the channel-coordinated feature set to form a multimodal feature matrix, integrating structured and unstructured features to form a comprehensive representation. A variational autoencoder is used to reduce the dimensionality of the multimodal feature matrix, generating a 128-dimensional latent space representation that preserves core information and reduces redundancy.

[0026] Cross-channel access sequence modeling and conversion funnel calculation: We construct cross-channel access sequences based on user behavior timestamps, recording the order and time intervals of user visits across different channels. We use LSTM networks to model user path patterns, capturing the temporal dependencies of user behavior and understanding user conversion paths across channels. We calculate conversion funnel metrics for each channel based on these path patterns, quantifying user conversion efficiency at each marketing stage and forming structured efficiency features.

[0027] Channel quality warning: Continuously monitor sentiment characteristics. If negative reviews exceed the preset threshold, the channel quality warning mark will be triggered to remind marketers to pay attention to and optimize related channels.

[0028] Channel Co-correlation Analysis and Feature Generation: A graph convolutional network is used to analyze overlapping user access patterns and construct a channel correlation matrix to represent the strength of inter-channel correlations. Combining the latent space representation with the channel correlation matrix, a spectral clustering algorithm is used to generate channel co-correlation groups, clustering channels with similar user behaviors and correlation patterns. Based on the co-correlation groups generated by spectral clustering, the conversion path complementarity coefficients between channels within the group are calculated to quantify the complementary relationship between channels in the user conversion path. Efficiency features, sentiment features, and conversion path complementarity coefficients are integrated to output a three-dimensional feature tensor as the final channel co-correlation feature set, providing comprehensive feature input for subsequent analysis.

[0029] This method effectively integrates structured and unstructured features to form a unified feature representation, improving model input quality. A variational autoencoder extracts key features and reduces dimensionality, improving computational efficiency and reducing information redundancy. Sentiment monitoring and early warning mechanisms can promptly identify channel quality issues and enhance the responsiveness of marketing campaigns. Quantifying inter-channel synergies through conversion path complementarity coefficients helps optimize marketing resource allocation and improve overall marketing effectiveness.

[0030] In some embodiments, a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network is constructed, wherein a channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, and a deep comprehensive feature set is generated, further comprising:

[0031] Construct a Gaussian process as a probabilistic surrogate model, initialize the hyperparameter search space, use random sampling to generate the initial hyperparameter combination, train the deep clustering model and calculate the silhouette coefficient and Davies-Bouldin index, update the Gaussian process surrogate model, calculate the acquisition function and select the next set of hyperparameter combinations, iterate and optimize until the convergence conditions are met, record the optimal hyperparameter combination, verify the performance of the optimal hyperparameter combination on the test set, and output the final hyperparameter configuration.

[0032] In this technical solution, the method aims to build a deep clustering model that incorporates autoencoders, variational autoencoders, and generative adversarial networks. Through multi-stage feature extraction and optimization, a deep, comprehensive feature set is generated. Furthermore, a Gaussian process surrogate model is used to optimize hyperparameters and improve the model's clustering performance, ultimately enabling efficient clustering and accurate analysis of marketing channel data.

[0033] Deep Clustering Model Construction: The channel-co-correlation feature set is first fed into an autoencoder, which consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional feature representation, while the decoder restores the low-dimensional features to a representation close to the input data. This allows the autoencoder to extract the core features of the data, remove noise and redundant information, and provide a concise and representative feature representation for subsequent processing. These core features are then fed into a variational autoencoder. Unlike traditional autoencoders, a variational autoencoder incorporates the concept of probabilistic generative modeling during the encoding process. Rather than encoding the input as a fixed latent variable, it encodes the latent variable as the distribution parameters of the latent variable. During decoding, latent variables are first sampled from this distribution before decoding. This enables the variational autoencoder to generate new data points in the latent space and extract the distributional features of the data, providing a more statistically meaningful feature representation for the generative adversarial network. These distributional features are then fed into a generative adversarial network (GAN), which consists of a generator and a discriminator. The generator attempts to generate realistic data samples to deceive the discriminator, while the discriminator strives to distinguish between generated samples and real samples. Through this adversarial training method, the generative adversarial network can further optimize the feature extraction process, making the generated features more discriminative and robust, and ultimately generating a deep comprehensive feature set.

[0034] Hyperparameter Optimization: A Gaussian process is constructed as a probabilistic proxy model. This is a nonparametric regression model based on probability that can predict the distribution of new data points based on existing data points. During hyperparameter optimization, it can predict the performance of other hyperparameter combinations based on existing hyperparameter combinations and their corresponding model performance metrics, thereby guiding the hyperparameter search. The hyperparameter search space is initialized, and the value ranges and types of each hyperparameter are determined. Random sampling is then used to generate initial hyperparameter combinations within the search space. These combinations serve as the first batch of hyperparameter candidates to be evaluated. A deep clustering model is trained using these initial hyperparameter combinations, and the model's silhouette coefficient and Davies-Bouldin index are calculated. The silhouette coefficient measures the ratio of a sample's similarity to its own cluster to its dissimilarity to other clusters; higher values ​​indicate better clustering. The Davies-Bouldin index evaluates clustering effectiveness by calculating the ratio of intra-cluster similarity to inter-cluster dissimilarity; lower values ​​indicate better clustering. These two metrics comprehensively reflect the quality of clustering and the degree of cluster separation. Based on the model training results, the Gaussian process surrogate model is updated to better fit the relationship between hyperparameters and model performance. Next, an acquisition function is calculated. Based on the surrogate model's predictions and uncertainty information, the acquisition function selects the next set of hyperparameter combinations most likely to improve performance. This training, evaluation, updating, and selection process is repeated, continuously optimizing the hyperparameter combination until convergence criteria are met, such as reaching the maximum number of iterations or the model performance improvement falling below a certain threshold. The optimal hyperparameter combination is recorded and its performance is verified on the test set to ensure good generalization. Finally, the validated optimal hyperparameter configuration is output, providing the optimal parameter settings for the deep clustering model.

[0035] By cascading autoencoders, variational autoencoders, and generative adversarial networks, this method can extract and optimize data features from different levels and perspectives. The resulting deep, comprehensive feature set is more representative and discriminative, helping to improve clustering accuracy and effectiveness. Using a Gaussian process surrogate model to assist in hyperparameter search can quickly locate potential high-quality hyperparameter regions, reduce unnecessary hyperparameter combination evaluations, and improve the efficiency and effectiveness of hyperparameter optimization, thereby providing more appropriate parameter configurations for deep clustering models. Using two complementary performance evaluation metrics, the silhouette coefficient and the Davies-Bouldin index, can comprehensively and objectively measure the quality of the clustering model, ensuring that the final clustering output performs well in terms of both intra-cluster compactness and inter-cluster separation.

[0036] In some embodiments, based on the deep integrated feature set, a clustering algorithm is used to classify marketing channels into different categories, generating preliminary classification results of the marketing channels, evaluating clustering quality using an evaluation index, and determining the optimal number of clusters, including:

[0037] Input the deep comprehensive feature set into the K-Means clustering algorithm, perform mean clustering, and generate preliminary channel category attribution results;

[0038] Calculate the silhouette coefficient of the clustering results, evaluate the closeness between the sample and the same category and the separation between the sample and other categories, and judge the internal consistency of the clustering effect;

[0039] Calculate the Davies-Bouldin index, evaluate the clustering quality by the ratio of intra-class dispersion to inter-class distance, and judge the stability and accuracy of the clustering results;

[0040] According to the comprehensive evaluation results of the silhouette coefficient and Davies-Bouldin index, the clustering parameters are adjusted, the clustering effect is optimized, and the final channel category attribution results are generated.

[0041] In the aforementioned technical solution, the method aims to classify marketing channels using a clustering algorithm based on a deep, comprehensive feature set. The method then uses an evaluation index to assess clustering quality and determine the optimal number of clusters, ultimately generating the final channel classification results. This analysis allows for a better understanding of the characteristics of different marketing channels, providing a basis for optimizing marketing strategies.

[0042] K-Means Clustering Algorithm Application: The deep comprehensive feature set is fed into the K-Means clustering algorithm to perform mean clustering and generate preliminary channel classification results. The K-Means algorithm iteratively optimizes the centroids of clusters to make samples within the same cluster as similar as possible and samples between different clusters as different as possible, thereby achieving a preliminary classification of marketing channels.

[0043] Silhouette Coefficient Calculation and Evaluation: Calculate the silhouette coefficient of the clustering results to assess the closeness of the sample to the same category and the separation from other categories, thereby determining the internal consistency of the clustering effect. The silhouette coefficient typically ranges from -1 to 1. Higher values ​​indicate closer proximity to samples of the same category and greater separation from samples of different categories, indicating better clustering results.

[0044] Davies-Bouldin index calculation and evaluation: Calculate the Davies-Bouldin index, evaluate the clustering quality by the ratio of intra-class dispersion to inter-class distance, and judge the stability and accuracy of the clustering results.

[0045] The lower the value of the Davies-Bouldin index, the better the clustering result, the higher the separation between clusters, and the higher the compactness within the clusters.

[0046] Comprehensive clustering effect evaluation and optimization: Based on the comprehensive evaluation results of the silhouette coefficient and Davies-Bouldin index, clustering parameters, such as the number of clusters (K value) in the K-Means algorithm, are adjusted to optimize the clustering effect and generate the final channel classification results. By continuously adjusting parameters and recalculating the evaluation index, the parameter configuration that achieves the best clustering effect is found.

[0047] The K-Means algorithm used in the above method can clearly classify marketing channels into different categories, helping marketers quickly understand the distribution and characteristics of these channels. The silhouette coefficient and Davies-Bouldin index provide quantitative evaluation methods that objectively assess clustering effectiveness and provide a basis for parameter adjustment. By adjusting clustering parameters based on comprehensive evaluation results, clustering results can be continuously optimized, improving classification accuracy and stability.

[0048] In some embodiments, based on the deep integrated feature set, a clustering algorithm is used to classify marketing channels into different categories, generating preliminary classification results of the marketing channels, evaluating clustering quality through an evaluation index, determining the optimal number of clusters, and outputting channel classification results, which then includes:

[0049] Analyze the customer retention rate and repurchase rate of each channel in the channel category classification results, and generate channel commercial value ranking results based on the conversion rate indicator;

[0050] Generate an allocation plan based on the channel commercial value ranking results based on the overlapping patterns of user behavior;

[0051] Track the difference in ROI after executing the resource allocation plan, and dynamically adjust the resource allocation plan to generate updated channel optimization directions.

[0052] In the above technical solution, the above method further analyzes the commercial value of each channel, generates a channel commercial value ranking result, and formulates a resource allocation plan based on user behavior patterns. By tracking the implementation effect of the resource allocation plan, resource allocation is dynamically adjusted to optimize the commercial value and return on investment of marketing channels, achieving efficient utilization and precise delivery of marketing resources.

[0053] Channel Commercial Value Analysis: This involves thoroughly analyzing channel classification results and assessing customer long-term value and loyalty based on customer retention and repurchase rates across each channel. Conversion rates are also considered to comprehensively reflect a channel's ability to acquire new customers and drive purchases. Based on these metrics, a comprehensive commercial value score is calculated for each channel and ranked. This helps identify high-value channels and provides a basis for resource allocation.

[0054] Resource Allocation Plan Generation: Rationally allocate marketing resources based on overlapping user behavior patterns. For channels with similar user groups or behavior patterns, consider collaborative marketing strategies to improve resource utilization efficiency.

[0055] Dynamic Optimization and Tracking Evaluation: After resource allocation plans are implemented, we closely monitor the ROI of each channel to identify high- and low-performing channels. Based on ROI feedback, we dynamically adjust resource allocation, optimize the channel mix, and improve overall marketing effectiveness.

[0056] This approach comprehensively considers key metrics such as customer retention, repurchase rate, and conversion rate to more comprehensively assess the commercial value of a channel and provide a scientific basis for resource allocation. Based on overlapping patterns in user behavior, it enables precise resource allocation, improving the synergy of marketing activities and resource utilization efficiency. By tracking return on investment and dynamically adjusting resource allocation, it ensures that marketing strategies adapt to market changes and continuously improves marketing effectiveness.

[0057] In some embodiments, analyzing the customer retention rate and repurchase rate of each channel in the channel category classification results and combining them with the conversion rate index to generate a channel commercial value ranking result includes:

[0058] Extract customer retention rate, repurchase rate and conversion rate indicators for each channel and use the weighted average method to combine customer retention rate, repurchase rate and conversion rate to calculate the comprehensive business value score of each channel;

[0059] According to the comprehensive commercial value score, each channel is ranked to generate the channel commercial value ranking results; based on the preset threshold, each channel is divided into high-value channels and low-value channels.

[0060] In the aforementioned technical solution, the method aims to calculate each channel's overall commercial value score by comprehensively evaluating its customer retention rate, repurchase rate, and conversion rate, and to rank the channels accordingly. Furthermore, channels are categorized as high-value or low-value based on preset thresholds, providing a basis for subsequent marketing resource allocation and strategy optimization.

[0061] Extract key metrics: Extract customer retention rate, repurchase rate, and conversion rate for each channel from the channel classification results. These metrics reflect customer loyalty within the channel, repeat purchase behavior, and the effectiveness of marketing activities, respectively.

[0062] Calculating a comprehensive business value score: We use a weighted average method to combine customer retention rate, repurchase rate, and conversion rate to calculate a comprehensive business value score for each channel. This step assigns weights to different indicators, comprehensively considering their contribution to business value, and deriving a score that reflects the channel's overall business value.

[0063] Generate channel commercial value ranking results: Rank channels based on their comprehensive commercial value scores to generate a channel commercial value ranking result. This ranking result intuitively demonstrates the relative commercial value of each channel, providing marketers with a clear channel priority.

[0064] Channel classification: Based on pre-set thresholds, channels are categorized as high-value or low-value. This step, by setting clear classification criteria, helps marketers quickly identify high-performing channels and those in need of improvement.

[0065] This approach comprehensively considers multiple metrics, including customer retention rate, repurchase rate, and conversion rate, to more comprehensively assess the commercial value of a channel, avoiding the one-sidedness of single-metric evaluation. This ranking and classification provides a clear basis for allocating marketing resources and formulating strategies, helping prioritize high-value channels and improve overall marketing effectiveness. Pre-set thresholds can be adjusted based on a company's specific goals and market conditions, making channel classification more tailored to its specific needs.

[0066] In some embodiments, based on the overlapping patterns of user behaviors, a distribution plan is generated for the channel commercial value ranking results, including:

[0067] Analyze the overlapping patterns of user behavior. If there are overlapping patterns, generate a cross-channel joint advertising plan through generative adversarial networks. If there are no overlapping patterns of user behavior in the channels, analyze their resource utilization efficiency through K-Means clustering and generate a resource reduction plan.

[0068] In this technical solution, the method generates resource allocation plans for different marketing channels based on overlapping user behavior patterns. For channels with overlapping user behavior, a generative adversarial network is used to create a joint advertising plan to enhance the relevance and effectiveness of the ads. For channels without overlapping user behavior, K-Means clustering is used to analyze their resource utilization efficiency and generate resource reduction plans to achieve efficient resource allocation.

[0069] Analysis of overlapping user behavior patterns: Collect user behavior data across different marketing channels, including browsing history, click behavior, and purchase behavior. Preprocess this data, including cleaning, noise removal, and missing value filling, to ensure data quality and usability. Data mining and analysis techniques are used to identify overlapping user behavior patterns across different channels. For example, this analysis can include whether users browsed the same or similar products across multiple channels, or whether they initially researched products on one channel before completing a purchase on another.

[0070] Cross-channel joint advertising plan generation: If there are overlapping patterns in user behavior, a generative adversarial network (GAN) is used to generate cross-channel joint advertising plans. GAN can generate new, creative advertising content and delivery strategies based on existing user behavior data. These plans are designed to increase the relevance and appeal of ads, thereby improving conversion rates and user engagement. By analyzing user behavior paths and preferences across different channels, the timing, frequency, and content of advertising are optimized to achieve precision marketing. For example, ads relevant to users' interests can be delivered on channels they frequently visit, and advertising strategies can be dynamically adjusted based on real-time user behavioral feedback.

[0071] Resource reduction plan generation: If channels lack overlapping user behavior patterns, the K-Means clustering algorithm is used to analyze their resource utilization efficiency. K-Means clustering can categorize channels based on metrics such as resource utilization efficiency, identifying those with low resource utilization efficiency. Based on the clustering results, a resource reduction plan is developed to shift resources from inefficient to more efficient channels, improving overall resource utilization efficiency. This may include reducing advertising budgets on inefficient channels and optimizing staffing.

[0072] By analyzing overlapping user behavior patterns, this method enables precise cross-channel advertising, improving advertising effectiveness and conversion rates. For channels without overlapping user behavior, K-Means clustering is used to analyze resource utilization efficiency, generate resource reduction plans, optimize resource allocation, and avoid resource waste. By flexibly adopting different strategies based on different user behavior patterns, marketing campaigns are more flexible and adaptable.

[0073] According to another aspect of the present invention, a heterogeneous data-driven marketing channel clustering modeling and optimization system is provided, based on the above method; the system comprises:

[0074] An acquisition module is used to acquire a multi-dimensional data set of marketing channels and perform preprocessing; the data set includes structured data and unstructured data;

[0075] A feature extraction module is used to extract text segmentation results and sentiment features from the unstructured data, and to extract image features to generate a structured feature set and an unstructured feature set;

[0076] An analysis module, configured to extract efficiency features based on the structured feature set, combine the sentiment features in the unstructured feature set, analyze overlapping patterns of cross-channel user behaviors, and generate a channel collaborative correlation feature set;

[0077] A deep clustering module is used to build a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network. The channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set.

[0078] The prediction module is used to classify marketing channels into different categories based on a deep comprehensive feature set and a clustering algorithm, generate preliminary classification results of marketing channels, evaluate the clustering quality through an evaluation index, and determine the optimal number of clusters to output the channel category attribution results.

[0079] In the above technical solution, in order to better use the above method, this application proposes a heterogeneous data-driven marketing channel clustering modeling and optimization system, each module corresponds to each step of the above method, and its specific principles have been described above and will not be repeated here.

[0080] According to another aspect of the present invention, a heterogeneous data-driven marketing channel clustering modeling and optimization device is provided, comprising:

[0081] at least one processor and a memory communicatively coupled to the at least one processor;

[0082] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above method.

[0083] In the above technical solution, in order to better run and process the method, the above method is stored in a memory and a processor is used to execute the stored method. It should be noted that the principle and effect of each step have been described above and will not be further explained here.

[0084] According to another aspect of the present invention, a computer-readable storage medium is provided, storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0085] In the above technical solution, in order to better run and use the method, the above method is stored in a computer-readable storage medium and implemented by a processor. It should be noted that the principle and effect of each step have been described above and will not be further explained here. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0087] Figure 1 This is a flow chart of an embodiment of a heterogeneous data-driven marketing channel clustering modeling and optimization method of the present invention;

[0088] Figure 2 This is a model diagram of an embodiment of a heterogeneous data-driven marketing channel clustering modeling and optimization method of the present invention;

[0089] Figure 3 It is a structural diagram of an embodiment of a heterogeneous data-driven marketing channel clustering modeling and optimization system of the present invention. DETAILED DESCRIPTION

[0090] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It is particularly noted that the following examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. Similarly, the following examples are only some embodiments of the present invention and are not intended to be exhaustive. All other embodiments obtained by those of ordinary skill in the art without creative effort are intended to fall within the scope of protection of the present invention.

[0091] Example 1

[0092] See also Figure 1 , a heterogeneous data-driven marketing channel clustering modeling and optimization method, the method comprising:

[0093] S1. Obtain a multi-dimensional data set of marketing channels and perform preprocessing; the data set includes structured data and unstructured data;

[0094] Exemplarily, a multi-dimensional data set of marketing channels is obtained, and the data set includes structured data and unstructured data, the structured data includes the average daily number of independent visitors and order conversion rate, and the unstructured data includes user comment text and advertising video thumbnails. In this embodiment, the data comes from multiple e-commerce and advertising platforms, covering the advertising performance of each channel within 90 days, and is used to analyze the characteristics of different channels and optimize marketing strategies. The data set contains 889 records and 13 key features, covering structured data (such as daily average UV, order conversion rate, stay time, bounce rate, user click-through rate and other quantitative indicators) and unstructured data (such as user comments, advertising material text, video thumbnails, etc.). Structured data is used to quantify the core operating indicators of each channel, while unstructured data is extracted through natural language processing (NLP) and computer vision methods to extract deep features, providing richer information for cluster analysis. For example, user comments can reflect the audience feedback of the advertisement, and the advertising material text can reveal the main purpose of the marketing content. These data help to identify the differences between channels and provide support for subsequent optimization strategies. Furthermore, the IQR method and Z-score were used to detect outliers of numerical features, and visual analysis was performed through box plots. The KNN interpolation method was used to fill in missing numerical data to process abnormal data. For text data with a high missing ratio, deletion or word vector filling strategy based on Word2Vec was adopted to standardize the numerical data using the Min-Max normalization method to ensure that data of different dimensions were on the same scale.

[0095] S2. Extracting text segmentation results and sentiment features from the unstructured data, and extracting image features to generate a structured feature set and an unstructured feature set;

[0096] For example, for unstructured data, text data (such as user comments and advertising copy) uses the BERT (Bidirectional Encoder Representations from Transformers) model to extract high-dimensional feature vectors, and uses principal component analysis (PCA) for dimensionality reduction to reduce data redundancy and improve computational efficiency. Image or video data (such as advertising materials) uses the ResNet pre-trained model to extract key visual features and map them to the same feature space as structured data for unified analysis. Ultimately, after data preprocessing and feature engineering, all data is converted into standardized feature representations that can be used for deep clustering analysis, providing high-quality input data for subsequent model training.

[0097] S3. Extracting efficiency features based on the structured feature set, combining them with the sentiment features in the unstructured feature set, analyzing overlapping patterns of cross-channel user behaviors to generate a channel collaborative correlation feature set;

[0098] In this embodiment, S3, extracting efficiency features based on the structured feature set, combining them with the sentiment features in the unstructured feature set, analyzing overlapping patterns of cross-channel user behaviors, and generating a channel collaborative correlation feature set, includes:

[0099] S31. Concatenate the channel collaborative correlation feature set to form a multimodal feature matrix, and use a variational autoencoder to reduce the dimensionality of the multimodal feature matrix to generate a 128-dimensional latent space representation;

[0100] For example, in marketing data processing, the channel collaborative correlation feature set includes a structured feature set (such as the user's click-through rate, conversion rate, purchase frequency, etc. in each channel) and an unstructured feature set (such as the feature vector after image feature extraction, text sentiment feature value, etc.). These feature sets are spliced ​​into a complete multimodal feature matrix. Since this matrix may have a very high dimension, in order to reduce the complexity and redundancy of the data, a variational autoencoder (VAE) is used for dimensionality reduction. The variational autoencoder maps the high-dimensional feature matrix to a 128-dimensional latent space representation by learning the potential distribution of the data, which can not only retain the key information in the data but also reduce the computational burden.

[0101] S32. Build cross-channel access sequences based on user behavior timestamps, use LSTM networks to model user path patterns, calculate conversion funnel indicators for each channel, and generate structured efficiency features.

[0102] For example, consider a user's behavior on an e-commerce platform. Within a single day, the user might first browse product recommendations on a social media platform (timestamp 09:30), then search for product details through a search engine (timestamp 09:45), and finally complete a purchase on the e-commerce platform's app (timestamp 10:15). Based on these timestamps, a cross-channel visit sequence is constructed: [Social Media\rightarrow Search Multi-Channel Visit Sequence: Social Media → Search Engine → E-commerce Platform App]. This sequence is modeled using a long short-term memory (LSTM) network, which captures the temporal dependencies and behavioral path patterns of users across different channels. By analyzing behavioral data such as user dwell time and bounce rate across channels, metrics such as conversion rate for each channel in the conversion funnel (e.g., exposure, click, purchase, etc.) are calculated to generate structured efficiency features. For example, a click-through conversion rate of 15% for social media channels and a purchase conversion rate of 10% for search engines can be used.

[0103] S33. If the sentiment feature shows that the negative evaluation exceeds the threshold, a channel quality warning mark is triggered;

[0104] For example, when performing sentiment analysis on user comments, suppose an e-commerce company collects a batch of user comments on a social media channel, and the sentiment analysis algorithm calculates that the proportion of negative comments in these comments exceeds a preset threshold (such as a threshold of 30%). At this time, the system automatically triggers a channel quality warning mark to remind operators that the channel may have problems with product quality, service experience, etc., which require timely attention and processing.

[0105] S34. Analyze user overlapping access patterns through graph convolutional networks and construct a channel correlation matrix;

[0106] For example, assume an e-commerce platform has multiple marketing channels, including social media, search engines, email, and offline stores. By analyzing the overlap between user visits across these channels, we find that, for example, a large number of users visit search engines after visiting social media, and some also visit offline stores to experience products. We use a graph convolutional network (GCN) to model this complex overlapping user visit relationship and construct a channel correlation matrix. In this matrix, the element value represents the strength of the correlation between two channels. For example, a correlation value of 0.7 between social media and search engines indicates a strong overlap between the two channels.

[0107] S35. Combining the latent space representation with the channel association matrix, a spectral clustering algorithm is used to generate channel collaborative groupings. Based on the grouping results, the conversion path complementarity coefficients between channels within the group are calculated, and a three-dimensional feature tensor containing efficiency features, sentiment features, and conversion path complementarity coefficients is output as the final channel collaborative association feature set.

[0108] For example, suppose we have a structured feature set containing 1 million user data, including user ID, number of purchases, browsing time (unit: minutes) and conversion status (0 for unconverted, 1 for converted). First, calculate the user conversion efficiency index using the formula: conversion efficiency = (number of purchases / browsing time) × 100 to obtain the conversion efficiency value of each user. For example, user A has 5 purchases and a browsing time of 50 minutes, so the conversion efficiency is (5 / 50) × 100 = 10.0. Then, combined with the sentiment tendency feature in the unstructured feature set, assume that the user comments are analyzed for sentiment using natural language processing technology to obtain a sentiment tendency score (ranging from -1 to 1). For example, the sentiment score of user A's comment is 0.8, indicating a positive sentiment. Next, we analyzed the overlapping patterns of cross-channel user behavior. Assuming that 30% of the user IDs in the user behavior data on e-commerce platforms and social media channels have active records in both channels, we calculated the average conversion efficiency of overlapping users (for example, 8.5 for e-commerce platforms and 7.2 for social media) and the average sentiment tendency (for example, 0.6 for e-commerce platforms and 0.5 for social media), and used a weighted average algorithm (weights 0.6 and 0.4, respectively) to derive a comprehensive indicator: the comprehensive conversion efficiency value is (8.5×0.6)+(7.2×0.4)=7.98, and the comprehensive sentiment tendency value is (0.6×0.6)+(0.5×0.4)=0.56. Finally, a channel collaboration feature set is generated. The above comprehensive indicators are combined with the user overlap ratio between channels (30%) to form a feature vector [7.98, 0.56, 0.3]. The feature vectors of all users are classified using a clustering algorithm (such as K-means, setting K = 3) to obtain feature groups of different channel collaboration patterns. For example, the mean of the feature vector of group 1 is close to [8.0, 0.6, 0.35], indicating a high-conversion and high-emotion collaboration pattern.

[0109] S4. Construct a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network. The channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set.

[0110] In this embodiment, S4, constructing a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network, inputting the channel collaborative correlation feature set into the autoencoder to extract core features; inputting the core features into the variational autoencoder to extract distribution features; and inputting the distribution features into the generative adversarial network to extract optimization features, generating a deep comprehensive feature set, further comprising:

[0111] S41. Construct a Gaussian process as a probabilistic proxy model, initialize the hyperparameter search space, use random sampling method to generate the initial hyperparameter combination, train the deep clustering model and calculate the silhouette coefficient and Davies-Bouldin index, update the Gaussian process proxy model, calculate the acquisition function and select the next set of hyperparameter combinations, iterate and optimize until the convergence conditions are met, record the optimal hyperparameter combination, verify the performance of the optimal hyperparameter combination on the test set, and output the final hyperparameter configuration.

[0112] The deep clustering model constructed in this application integrates three deep learning architectures: AutoEncoder, Variational AutoEncoder (VAE) and Generative Adversarial Network (GAN), and adopts a layered processing mechanism to fully explore the characteristics of heterogeneous marketing data and optimize the clustering effect. Figure 2The autoencoder is fed with a set of channel-related features from an e-commerce company. The encoder compresses the high-dimensional feature data into a low-dimensional core feature representation, removing redundant information and retaining the most important features. The decoder then restores the core features to data similar to the original input. By training the autoencoder, it accurately reconstructs the input data and extracts effective core features. For example, a feature set containing multi-dimensional information such as user behavior and sentiment across different marketing channels is fed into the autoencoder. After training, the autoencoder can extract key core features that reflect user behavior patterns and channel associations. The extracted core features are then fed into a variational autoencoder (VAE). Unlike conventional autoencoders, the VAE encoder outputs a probability distribution (mean and variance vector) rather than a fixed vector. Using reparameterization techniques, latent variables are sampled from this probability distribution, which are then decoded by the decoder to generate data. In e-commerce marketing scenarios, VAEs can learn the latent distribution of core features and extract distributional features. These distributional features can better capture the uncertainty in the data and the correlation between different features, providing richer information for subsequent clustering analysis. For example, VAEs can reveal the underlying distribution patterns of user behavior characteristics under different marketing channel combinations, as well as the distribution of associations between different emotional profiles and channel behavior characteristics. The distribution features extracted by VAEs are then fed into a generative adversarial network (GAN), which consists of a generator and a discriminator. The generator's goal is to generate data that is as realistic as possible, making it impossible for the discriminator to distinguish between generated data and real data. The discriminator, on the other hand, strives to distinguish between real and generated data. During this adversarial training process, the generator continuously learns how to generate samples that are closer to real data, thereby optimizing the data's features. In the task of marketing channel clustering, GANs can further optimize features, making the generated optimized features more prominently highlight the differences between different channel combinations and user behavior patterns, thereby improving clustering accuracy. For example, the optimized features generated by the generator can more clearly reflect which channel combinations are more likely to lead users to purchase, as well as the relationship between user behavior patterns under different emotional tendencies and channel selection. Ultimately, a deep and comprehensive feature set is generated. This comprehensive feature set combines the core features extracted by the autoencoder, the distribution features extracted by VAE, and the features optimized by GAN. It can more comprehensively and deeply characterize the collaborative relationship between marketing channels and user behavior characteristics, providing high-quality feature representation for subsequent clustering analysis.

[0113] During the training process, an unsupervised learning paradigm is adopted, and the reconstruction error of the autoencoder is combined to optimize feature learning. At the same time, the K-Means objective function is jointly optimized to ensure that the learned features can directly serve the clustering task. The model parameters are adjusted using the Bayesian optimization algorithm, and the training process is optimized by the mini-batch gradient descent (SGD) algorithm to improve the convergence speed and accuracy of the model. The core innovations of this model are: (1) improving the data reconstruction quality through the feature learning mechanism of the autoencoder, making the feature representation more stable; (2) using VAE to model the latent space, making the data distribution more continuous and robust; (3) generating data samples close to the real distribution through GAN, enhancing the model's adaptability to complex marketing data; (4) adopting a joint optimization strategy to deeply integrate feature learning with clustering objectives, forming an end-to-end optimization framework, thereby improving clustering effect and model generalization ability.

[0114] The training and optimization process aims to improve clustering effectiveness and optimize marketing channel classification through deep feature learning mechanisms using autoencoders, variational autoencoders (VAEs), and generative adversarial networks (GANs). The following describes several key points in model construction.

[0115] (1) Autoencoder feature learning

[0116] Through its encoder-decoder architecture, autoencoders are able to learn the essential characteristics of input data. During the encoding phase, the network compresses high-dimensional data into a low-dimensional latent code; during the decoding phase, the network attempts to reconstruct the original data from the latent code. By minimizing the reconstruction error, the network is forced to retain the most discriminative features in the data. This mechanism is particularly effective in feature extraction from marketing data, capturing user behavior patterns across different marketing channels. Through this process, the model can provide high-quality feature representations for subsequent clustering tasks.

[0117] (2) Probabilistic Modeling of Variational Autoencoders

[0118] Building on the standard autoencoder, the variational autoencoder (VAE) introduces the concept of probabilistic generation by constraining the latent space to conform to a standard normal distribution. This modeling approach enhances robustness to data noise, enabling the extraction of valuable information from cluttered data. It also improves the continuity of the feature space, making the generated latent features smoother and more stable. Finally, it provides the ability to generate synthetic samples with marketing significance. Especially when processing cross-platform data, VAE can effectively reduce differences in data distribution, thereby improving the model's adaptability and stability.

[0119] (3) Generative Adversarial Training Mechanism

[0120] Generative adversarial networks (GANs) explore the boundaries of data distribution through a game of generators and discriminators. The generator continuously generates new channel feature samples, while the discriminator learns to distinguish between real samples and generated samples. Through this adversarial process, GANs can effectively improve the model's adaptability to emerging channel types and enhance its ability to capture complex data patterns. GANs also provide strong support for data augmentation, especially when data samples are insufficient or imbalanced, effectively generating realistic synthetic samples and improving training results.

[0121] (4) Joint optimization strategy

[0122] The proposed deep clustering model employs a joint optimization strategy, simultaneously optimizing the K-Means clustering objective during deep feature learning. Specifically, the network not only learns feature representations but also dynamically updates the locations of cluster centers during training, gradually forming a distribution in the feature space suitable for clustering. This joint optimization strategy overcomes the feature drift problem inherent in traditional two-stage methods, thereby improving the stability and accuracy of clustering results.

[0123] To further improve model performance, a Bayesian optimization framework was used to automatically search for model hyperparameters. Bayesian optimization constructs a probabilistic surrogate model of the objective function to intelligently select the most promising parameter combinations for experimentation. Compared to traditional grid search and random search methods, Bayesian optimization can find more optimal hyperparameter configurations using the same computing resources, significantly improving model training efficiency and generalization. Specifically, a Gaussian process was constructed as a probabilistic surrogate model for hyperparameter optimization. During the hyperparameter search for a deep clustering model, the Bayesian optimization framework was used to determine the optimal parameter combination, and the probabilistic surrogate model was used for efficient search. First, a set of hyperparameter combinations was initialized, such as a learning rate of 0.001, 256 hidden layer neurons, and 10 cluster centers. Ten initial parameter combinations were generated using random sampling. The performance of the deep clustering model was evaluated by training it on the dataset, and the loss for each combination was recorded. For example, the loss for initial combination 1 was 2.35, while the loss for combination 2 was 2.48. Next, a probabilistic surrogate model was constructed, using Gaussian process regression as the surrogate model. The loss values ​​of the initial parameter combinations were used as training data. A mapping relationship between the parameter space and the loss values ​​was fitted, and the posterior probability distribution of each parameter combination was calculated. The mean and variance of the loss for untried parameter combinations were predicted. For example, the loss for a combination with a learning rate of 0.002 and 512 neurons was predicted to be 2.10 with a variance of 0.05. An acquisition function (e.g., expected improvement, EI) was then used to select the next set of parameter combinations to be evaluated. The parameter combination with the highest EI value, calculated to be 0.78, was a learning rate of 0.0025, a number of neurons of 384, and 12 cluster centers, indicating the greatest potential for improvement. This combination was then trained on a deep clustering model, resulting in an actual loss of 2.08. The surrogate model dataset was then updated. After 10 rounds of this process, the combination with a learning rate of 0.0022, 400 neurons, and 11 cluster centers achieved the lowest loss of 1.95, thus being identified as the optimal parameter combination.

[0124] S5. Based on the deep comprehensive feature set, the marketing channels are divided into different categories using clustering algorithms to generate preliminary classification results of the marketing channels. The clustering quality is evaluated through the evaluation index and the optimal number of clusters is determined to output the channel category attribution results.

[0125] In this embodiment, S5, based on the deep comprehensive feature set, uses a clustering algorithm to classify marketing channels into different categories, generates preliminary classification results of marketing channels, evaluates clustering quality through an evaluation index, and determines the optimal number of clusters, including:

[0126] S51, input the deep comprehensive feature set into the K-Means clustering algorithm, perform mean clustering, and generate preliminary channel category attribution results;

[0127] S52. Calculate the silhouette coefficient of the clustering results to evaluate the closeness between the sample and the same category and the separation between the sample and other categories, and determine the internal consistency of the clustering effect;

[0128] S53. Calculate the Davies-Bouldin index to evaluate clustering quality by the ratio of intra-class dispersion to inter-class distance, and judge the stability and accuracy of the clustering results.

[0129] S54. Based on the comprehensive evaluation results of the silhouette coefficient and the Davies-Bouldin index, adjust the clustering parameters, optimize the clustering effect, and generate the final channel category attribution result.

[0130] For example, an e-commerce company inputs a deep comprehensive feature set into the K-Means clustering algorithm. This deep comprehensive feature set integrates multi-dimensional information of marketing channels, including user behavior characteristics, emotional characteristics, conversion path characteristics, etc. Suppose there are 1,000 marketing channel samples, each of which is composed of the above features. Set the number of cluster centers of K-Means to 5, and the algorithm starts to perform mean clustering. By calculating the Euclidean distance between each sample and the cluster center, the sample is assigned to the category to which the nearest cluster center belongs. After several iterations, the cluster center is continuously updated, and finally converges to obtain a preliminary channel category attribution result, dividing the 1,000 marketing channel samples into 5 preliminary categories, each category representing a group of marketing channels with similar comprehensive characteristics.

[0131] Calculate the silhouette coefficient of the clustering results. The silhouette coefficient ranges from [-1, 1], with larger values ​​indicating better clustering results. Specifically, for each sample, calculate its average distance from other samples in the same category (the a value, indicating how closely the sample is aligned with the category to which it belongs), as well as its average distance from all samples in the nearest other category (the b value, indicating how closely the sample is separated from the nearest other category). Silhouette coefficient s = (ba) / max (a, b). Assuming that the average distance a of a marketing channel sample from samples in the same category is 0.3, and the average distance b from the nearest other category is 0.7, then the silhouette coefficient s = (0.7-0.3) / max (0.3, 0.7) = 0.4. Averaging the silhouette coefficients of all samples yields an average silhouette coefficient of 0.6, indicating that the clustering results are well-consistent internally, with samples closely aligned with the same category and separated from different categories.

[0132] Calculate the Davies-Bouldin index. This index measures clustering quality by the ratio of the intra-class dispersion (the average distance from each class's samples to the cluster center) to the inter-class distance (the distance between different cluster centers). For each class i, calculate the intra-class dispersion di and the inter-class distance mj between class i and class j. For each pair of classes i and j, calculate (di + dj) / mj, then take the maximum value across all class pairs, and finally average all classes to obtain the Davies-Bouldin index. Suppose that after calculation, the Davies-Bouldin index is 0.8. Generally speaking, the smaller the index value, the more stable and accurate the clustering results. A value of 0.8 indicates that the clustering results are at a moderate level in terms of stability and accuracy, and there is still room for improvement.

[0133] Based on the comprehensive evaluation results of a Silhouette Coefficient of 0.6 and a Davies-Bouldin Index of 0.8, the current number of clusters, 5, was found to be acceptable but not optimal. Clustering parameters were adjusted, such as by experimenting with different values ​​for the number of cluster centers, such as 4 and 6. K-Means clustering was then re-run, and the Silhouette Coefficient and Davies-Bouldin Index were recalculated. After repeated adjustments and evaluations, it was found that when the number of cluster centers was 4, the average Silhouette Coefficient increased to 0.7, while the Davies-Bouldin Index decreased to 0.6. This indicates that the clustering effect has been optimized, with samples becoming more closely aligned with their categories, better separation between categories, and improved stability of the clustering results. The final classification of marketing channels was determined to be based on four cluster centers. Each channel was assigned to one of four categories, each representing a combination of marketing channels with similar comprehensive characteristics. This provides a clear basis for channel classification in the subsequent development of precision marketing strategies.

[0134] In this embodiment, S5, based on the deep comprehensive feature set, a clustering algorithm is used to classify marketing channels into different categories, generating preliminary classification results of marketing channels, evaluating clustering quality through an evaluation index, determining the optimal number of clusters, and outputting channel category attribution results, which then includes:

[0135] S6. Analyze the customer retention rate and repurchase rate of each channel in the channel category classification results, and generate a channel commercial value ranking result based on the conversion rate indicator;

[0136] In this embodiment, S6, analyzing the customer retention rate and repurchase rate of each channel in the channel category classification results, and combining the conversion rate index to generate a channel commercial value ranking result, including:

[0137] S61. Extract the customer retention rate, repurchase rate, and conversion rate indicators for each channel and use the weighted average method to calculate the comprehensive business value score of each channel by combining the customer retention rate, repurchase rate, and conversion rate;

[0138] S62. Sort the channels according to the comprehensive commercial value score to generate a channel commercial value ranking result; based on a preset threshold, divide the channels into high-value channels and low-value channels.

[0139] Based on cluster analysis results, combined with key marketing metrics such as customer retention rate, repurchase rate, and order conversion rate, the value of different marketing channels is assessed. For example, cluster analysis reveals that channels in categories 0 and 3 perform exceptionally well in terms of marketing effectiveness and are therefore considered high-value channels, while channel category 2 is relatively inefficient. When optimizing marketing strategies, it's important not only to focus on improving individual channels but also to fully consider the synergies between channels. By combining the strengths of social media advertising and search engine advertising, it's possible to design collaborative advertising strategies that leverage the interactivity of social platforms and the precision of search engines to achieve optimal advertising effectiveness. This marketing channel optimization strategy is based on deep cluster analysis. Using deep learning models such as autoencoders, variational autoencoders (VAEs), and generative adversarial networks (GANs), the strategy combines structured and unstructured data to identify and segment underlying patterns within each channel. Cluster analysis reveals significant differences between channels. For example, some channels are high-value due to their high visit depth and order conversion rates, while others perform less effectively. Based on clustering results, refined management can be implemented for high-value channels, such as increasing advertising and strengthening customer service, further improving the marketing effectiveness of these channels. For low-value channels, cluster analysis reveals weaknesses, such as low user retention and conversion rates. Companies can optimize or gradually abandon these channels to avoid wasted resources and optimize overall marketing resource allocation. Considering the synergies between channels, companies can design joint advertising strategies to enhance their complementarity. For example, combining social media advertising with search engine advertising can maximize the advertising effectiveness of each channel.

[0140] Exemplarily, the customer retention rate, repurchase rate and conversion rate indicators of each channel are extracted and the weighted average method is adopted to calculate the comprehensive commercial value score of each channel by combining the customer retention rate, repurchase rate and conversion rate. According to the comprehensive commercial value score, the channels are ranked to generate the channel commercial value ranking results. If the comprehensive commercial value score of a channel is higher than the preset threshold, it is marked as a high-value channel. If the comprehensive commercial value score of a channel is lower than the preset threshold, it is marked as a low-value channel. According to the channel commercial value ranking results, channel optimization strategy recommendations are generated, including resource allocation for high-value channels and adjustment plans for low-value channels. Specifically, when analyzing the customer retention and repurchase rates of each channel in the channel classification results and combining them with conversion rate metrics to generate the channel commercial value ranking results, we first collect and clean data to obtain customer behavior data for each channel. For example, let's assume there are three channels, A, B, and C, with a total customer base of 10,000, 8,000, and 5,000, respectively. The retention rates for Channel A are 50%, Channel B is 40%, and Channel C is 30%. The repurchase rates for Channel A are 35%, Channel B is 25%, and Channel C is 20%. The conversion rates for Channel A are 10%, Channel B is 8%, and Channel C is 5%. Next, we use an algorithm to calculate the comprehensive indicators for each channel, using a weighted scoring method, setting a weight of 0.4 for retention rate, 0.3 for repurchase rate, and 0.3 for conversion rate. The calculation formula is: Comprehensive Score = Retention Rate × 0.4 + Repurchase Rate × 0.3 + Conversion Rate × 0.3. Taking Channel A as an example, its combined score is 0.5 × 0.4 + 0.35 × 0.3 + 0.1 × 0.3 = 0.2 + 0.105 + 0.03 = 0.335; Channel B's score is 0.4 × 0.4 + 0.25 × 0.3 + 0.08 × 0.3 = 0.16 + 0.075 + 0.024 = 0.259; and Channel C's score is 0.3 × 0.4 + 0.2 × 0.3 + 0.05 × 0.3 = 0.12 + 0.06 + 0.015 = 0.195. Calculations show that Channel A has the highest score, followed by Channel B, and finally Channel C. Further analysis shows that Channel A significantly outperforms other channels in retention and repurchase rates, indicating strong customer stickiness. Combined with conversion rate data, this suggests it has the highest commercial value. Channel C, on the other hand, performs poorly on all metrics and requires operational strategy optimization. Subsequently, the comprehensive score was combined with the channel input cost. Assuming that the costs of channels A, B, and C were 1 million, 800,000, and 500,000 respectively, the unit cost scores were calculated. Channel A was 0.335 / 100=0.00335, Channel B was 0.259 / 80=0.00324, and Channel C was 0.195 / 50=0.0039. It was found that channel C had the highest unit cost score, suggesting that there is still room for optimization at low cost.Finally, the business value ranking results are generated, and the comprehensive score ranking is A>B>C. However, combined with the unit cost score, it is recommended to prioritize the optimization of channel C to improve overall benefits. All calculations and rankings are processed automatically by the system, and the data is updated in real time to ensure accurate analysis.

[0141] S7. Generate an allocation plan based on the channel commercial value ranking results based on the overlapping patterns of user behaviors;

[0142] In this embodiment, S7, based on the overlapping patterns of user behaviors, generates an allocation plan for the channel commercial value ranking results, including:

[0143] S71. Analyze overlapping patterns of user behavior. If there are overlapping patterns, generate a cross-channel joint advertising plan through a generative adversarial network. If there are no overlapping patterns of user behavior in the channels, analyze their resource utilization efficiency through K-Means clustering to generate a resource reduction plan.

[0144] For example, we analyze overlapping patterns in user behavior. If there are overlapping patterns, we generate a cross-channel joint advertising plan through a generative adversarial network. We combine the synergy between social media advertising and search engine advertising to design a linkage advertising strategy. If there are no overlapping patterns in user behavior in low-value channels, we analyze their resource utilization efficiency through K-Means clustering and generate a resource reduction plan. We combine the channel classification results and key marketing indicators to dynamically adjust the advertising budget and optimize resource allocation. Based on the clustering results of the deep clustering model, we evaluate the effectiveness of the cross-channel joint advertising plan. We verify the effectiveness of the strategy through ROI analysis and optimize marketing decisions based on the customer lifetime value. Specifically:

[0145] When analyzing channel value and user behavior patterns, the system first uses the data collection module to obtain user behavior data from each channel. For example, it collects click-through rate and conversion rate data from 1 million users in high-value channels A and B over the past 30 days. It finds that Channel A has a click-through rate of 5.2% and a conversion rate of 1.8%, while Channel B has a click-through rate of 4.8% and a conversion rate of 1.5%. Using the Pearson correlation coefficient, the user behavior overlap between the two channels is 0.75, indicating significant overlap. Based on this, the system automatically triggers the cross-channel joint advertising plan generation module. Using historical data, it predicts the effectiveness of joint advertising. Assuming a single-channel advertising cost of 10 yuan per thousand impressions, joint advertising can reduce costs by 15%, to 8.5 yuan per thousand impressions, while also increasing conversion rates to 2.0%. A linear regression model predicts an increase of 5,000 conversions over 30 days after the joint advertising campaign. The system then generates a budget allocation plan, with Channel A receiving 60% of the budget, or 1.8 million yuan, and Channel B receiving 40%, or 1.2 million yuan, to ensure optimal resource allocation. The system then analyzed user behavior data for low-value channels C and D. Over the past 30 days, they had 500,000 and 400,000 users, respectively, with click-through rates of only 2.1% and 1.9%, and conversion rates as low as 0.5% and 0.4%. Cluster analysis revealed no significant correlation between user behavior across the two channels, with a correlation coefficient of only 0.2, below the threshold of 0.5. This automatically triggered the resource reduction plan generation module. Using the channel contribution scoring algorithm, the system calculated the contribution of channels C and D to be 3 and 2 out of 10, respectively, below the minimum threshold of 5. Therefore, it recommended a 50% reduction in the advertising budget, from 1 million yuan to 500,000 yuan. The saved resources would be reallocated to a joint campaign plan with high-value channels, forming a closed-loop resource optimization logic. To ensure that low-value channels maintained a minimum level of exposure, the system implemented a minimum budget protection mechanism, ensuring at least 200,000 yuan in operating expenses. Furthermore, the system tracked the effectiveness of the reductions through a real-time monitoring module. If the conversion rate increased above 0.8%, the resource allocation strategy would be reassessed to ensure the scientific nature of the dynamic adjustments.

[0146] S8. Track the difference in return on investment after the resource allocation plan is implemented, and dynamically adjust the resource allocation plan to generate an updated channel optimization direction.

[0147] For example, in the process of tracking the difference in return on investment after the implementation of the resource allocation plan and dynamically adjusting the resource allocation plan to generate an updated channel optimization direction, information technology can be used to achieve full process automation. First, assume that a company's initial resource allocation plan is: online channel A invests 500,000 yuan, offline channel B invests 300,000 yuan, and online channel C invests 200,000 yuan, with expected returns on investment (ROI) of 15%, 10%, and 12%, respectively. The actual ROI data of each channel is automatically collected through the data collection system. Assume that after one month of implementation, the actual ROI of channel A is 13.5%, channel B is 11.2%, and channel C is 9.8%. Using an algorithm to analyze the differences, the calculation formula is: difference value = actual ROI - expected ROI, and the difference for channel A is -1.5%, channel B is +1.2%, and channel C is -2.2%. The system adjusts weights based on the variance, using a weighted adjustment algorithm: New Allocation Ratio = Original Ratio × (1 + Variance Value / Expected ROI). The calculated new ratio for Channel A is 50 × (1-1.5 / 15) = 475,000 RMB, for Channel B it is 30 × (1+1.2 / 10) = 336,000 RMB, and for Channel C it is 20 × (1-2.2 / 12) = 169,000 RMB. Next, the system analyzes the correlation between each channel's conversion rate and investment using historical data. Assuming that Channel B's conversion rate increases linearly with investment (correlation coefficient 0.85), while Channel C's growth slows (correlation coefficient 0.6), and within the budget constraint of 1 million RMB, the optimized allocation is further fine-tuned to Channel A: 470,000 RMB, Channel B: 350,000 RMB, and Channel C: 180,000 RMB. Finally, the system generates an optimization direction report, indicating that resources should be tilted towards Channel B, while monitoring the marginal benefits of Channel C. If the conversion rate continues to be lower than 8%, it triggers the secondary adjustment logic, links other businesses such as promotional activity data, analyzes their impact on Channel C, forms a closed-loop optimization, and ensures that resource allocation matches market feedback in real time.

[0148] Example 2

[0149] See also Figure 3 A heterogeneous data-driven marketing channel clustering modeling and optimization system is based on the above method; the system includes:

[0150] An acquisition module is used to acquire a multi-dimensional data set of marketing channels and perform preprocessing; the data set includes structured data and unstructured data;

[0151] A feature extraction module is used to extract text segmentation results and sentiment features from the unstructured data, and to extract image features to generate a structured feature set and an unstructured feature set;

[0152] An analysis module, configured to extract efficiency features based on the structured feature set, combine the sentiment features in the unstructured feature set, analyze overlapping patterns of cross-channel user behaviors, and generate a channel collaborative correlation feature set;

[0153] A deep clustering module is used to build a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network. The channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set.

[0154] The prediction module is used to classify marketing channels into different categories based on a deep comprehensive feature set and a clustering algorithm, generate preliminary classification results of marketing channels, evaluate the clustering quality through an evaluation index, and determine the optimal number of clusters to output the channel category attribution results.

[0155] In the above technical solution, in order to better use the above method, this application proposes a heterogeneous data-driven marketing channel clustering modeling and optimization system, each module corresponds to each step of the above method, and its specific principles have been described above and will not be repeated here.

[0156] According to another aspect of the present invention, a heterogeneous data-driven marketing channel clustering modeling and optimization device is provided, comprising:

[0157] at least one processor and a memory communicatively coupled to the at least one processor;

[0158] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above method.

[0159] In the above technical solution, in order to better run and process the method, the above method is stored in a memory and a processor is used to execute the stored method. It should be noted that the principle and effect of each step have been described above and will not be further explained here.

[0160] According to another aspect of the present invention, a computer-readable storage medium is provided, storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0161] In the above technical solution, in order to better run and use the method, the above method is stored in a computer-readable storage medium and implemented by a processor. It should be noted that the principle and effect of each step have been described above and will not be further explained here.

[0162] The above descriptions are only some embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A heterogeneous data-driven marketing channel clustering modeling and optimization method, characterized by: The method comprises: Acquire a multi-dimensional data set of marketing channels and perform preprocessing; the data set includes structured data and unstructured data; Extracting text segmentation results and sentiment features from the unstructured data, and extracting image features to generate a structured feature set and an unstructured feature set; Extracting efficiency features based on the structured feature set, combining them with sentiment features in the unstructured feature set, analyzing overlapping patterns of cross-channel user behaviors to generate a channel collaborative correlation feature set; Construct a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network. The channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set. Based on a deep comprehensive feature set, a clustering algorithm is used to divide marketing channels into different categories, generate preliminary classification results of marketing channels, evaluate the clustering quality through an evaluation index, and determine the optimal number of clusters to output the channel category attribution results.

2. The heterogeneous data-driven marketing channel clustering modeling and optimization method according to claim 1, characterized in that: Based on the efficiency feature extracted from the structured feature set, combined with the sentiment feature in the unstructured feature set, cross-channel user behavior overlapping patterns are analyzed to generate a channel collaborative correlation feature set, including: Based on the efficiency feature extracted from the structured feature set, combined with the sentiment feature in the unstructured feature set, cross-channel user behavior overlapping patterns are analyzed to generate a channel collaborative correlation feature set, including: The channel collaborative correlation feature set is spliced ​​to form a multimodal feature matrix, and a variational autoencoder is used to reduce the dimensionality of the multimodal feature matrix to generate a 128-dimensional latent space representation; Build cross-channel access sequences based on user behavior timestamps, use LSTM networks to model user path patterns, calculate conversion funnel indicators for each channel, and generate structured efficiency features; If the sentiment feature shows that negative reviews exceed the threshold, a channel quality warning mark will be triggered; Analyze user overlapping access patterns through graph convolutional networks and construct a channel correlation matrix; Combining the latent space representation with the channel association matrix, a spectral clustering algorithm is used to generate channel co-groupings; Based on the grouping results, the conversion path complementarity coefficients between channels in the group are calculated, and a three-dimensional feature tensor containing efficiency features, sentiment features and conversion path complementarity coefficients is output as the final channel collaborative correlation feature set.

3. The heterogeneous data-driven marketing channel clustering modeling and optimization method according to claim 1, characterized in that: Constructing a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network, wherein the channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set, further comprising: Construct a Gaussian process as a probabilistic surrogate model, initialize the hyperparameter search space, use random sampling to generate the initial hyperparameter combination, train the deep clustering model and calculate the silhouette coefficient and Davies-Bouldin index, update the Gaussian process surrogate model, calculate the acquisition function and select the next set of hyperparameter combinations, iterate and optimize until the convergence conditions are met, record the optimal hyperparameter combination, verify the performance of the optimal hyperparameter combination on the test set, and output the final hyperparameter configuration.

4. The heterogeneous data-driven marketing channel clustering modeling and optimization method according to claim 1, characterized in that: Based on a deep comprehensive feature set, a clustering algorithm is used to classify marketing channels into different categories, generating preliminary classification results for marketing channels. The clustering quality is evaluated through an evaluation index to determine the optimal number of clusters, including: Input the deep comprehensive feature set into the K-Means clustering algorithm, perform mean clustering, and generate preliminary channel category attribution results; Calculate the silhouette coefficient of the clustering results, evaluate the closeness between the sample and the same category and the separation between the sample and other categories, and judge the internal consistency of the clustering effect; Calculate the Davies-Bouldin index, evaluate the clustering quality by the ratio of intra-class dispersion to inter-class distance, and judge the stability and accuracy of the clustering results; According to the comprehensive evaluation results of the silhouette coefficient and Davies-Bouldin index, the clustering parameters are adjusted, the clustering effect is optimized, and the final channel category attribution results are generated.

5. The heterogeneous data-driven marketing channel clustering modeling and optimization method according to claim 1, characterized in that: Based on a deep comprehensive feature set, a clustering algorithm is used to classify marketing channels into different categories, generating preliminary classification results for marketing channels. The clustering quality is evaluated through an evaluation index, and the optimal number of clusters is determined to output the channel category classification results. This includes: Analyze the customer retention rate and repurchase rate of each channel in the channel category classification results, and generate channel commercial value ranking results based on the conversion rate indicator; Generate an allocation plan based on the channel commercial value ranking results based on the overlapping patterns of user behavior; Track the difference in ROI after executing the resource allocation plan, and dynamically adjust the resource allocation plan to generate updated channel optimization directions.

6. The heterogeneous data-driven marketing channel clustering modeling and optimization method according to claim 5, characterized in that: Analyze the customer retention rate and repurchase rate of each channel in the channel category classification results, and combine them with the conversion rate indicator to generate the channel commercial value ranking results, including: Extract customer retention rate, repurchase rate and conversion rate indicators for each channel and use the weighted average method to combine customer retention rate, repurchase rate and conversion rate to calculate the comprehensive business value score of each channel; According to the comprehensive commercial value score, each channel is ranked to generate the channel commercial value ranking results; based on the preset threshold, each channel is divided into high-value channels and low-value channels.

7. The heterogeneous data-driven marketing channel clustering modeling and optimization method according to claim 5, characterized in that: Based on the overlapping patterns of user behavior, a distribution plan is generated for the channel commercial value ranking results, including Analyze the overlapping patterns of user behavior. If there are overlapping patterns, generate a cross-channel joint advertising plan through generative adversarial networks. If there are no overlapping patterns of user behavior in the channels, analyze their resource utilization efficiency through K-Means clustering and generate a resource reduction plan.

8. A heterogeneous data-driven marketing channel clustering modeling and optimization system, characterized by: Based on the method according to any one of claims 1 to 7; the system comprises: An acquisition module is used to acquire a multi-dimensional data set of marketing channels and perform preprocessing; the data set includes structured data and unstructured data; A feature extraction module is used to extract text segmentation results and sentiment features from the unstructured data, and to extract image features to generate a structured feature set and an unstructured feature set; An analysis module, configured to extract efficiency features based on the structured feature set, combine the sentiment features in the unstructured feature set, analyze overlapping patterns of cross-channel user behaviors, and generate a channel collaborative correlation feature set; A deep clustering module is used to build a deep clustering model including an autoencoder, a variational autoencoder, and a generative adversarial network. The channel collaborative correlation feature set is input into the autoencoder to extract core features; the core features are input into the variational autoencoder to extract distribution features; the distribution features are input into the generative adversarial network to extract optimization features, thereby generating a deep comprehensive feature set. The prediction module is used to classify marketing channels into different categories based on a deep comprehensive feature set and a clustering algorithm, generate preliminary classification results of marketing channels, evaluate the clustering quality through an evaluation index, and determine the optimal number of clusters to output the channel category attribution results.

9. A heterogeneous data-driven marketing channel clustering modeling and optimization device, characterized by: include: at least one processor and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images